Deep Learning for Human-Centered Semi-Autonomous Vehicles

• 반 자동 주행 차량을 위해서 운전자에게 초점을 맞출 필요가 있음
  • 왜?

1. Body Pose

1.1 목적

사고나 났을경우 올바른 자세(안전띠, 전방 주시)가 아니면 큰 피해로 연결됨

What Do We Need to Detect?

• Detection task 1: Hands on wheel
• Detection task 2: Position of head relative to head rest
• Detection task 3: Full upper body pose

1.2 연구 방법

• Sequential Detection Approach
  • Charles, James, et al. "Upper body pose estimation with temporal sequential forests." Proceedings of the British Machine Vision Conference 2014. BMVA Press, 2014.
• Temporal Convolutional Neural Networks
  • Pfister, Tomas, James Charles, and Andrew Zisserman. "Flowing convnets for human pose estimation in videos." Proceedings of the IEEE International Conference on Computer Vision. 2015.

2. Gaze Classification vs Gaze Estimation

Gaze Classification Pipeline

• Face detection (the only easy step)
• Face alignment (active appearance models or deep nets)
• Eye/pupil detection (are the eyes visible?)
• Head (and eye) pose estimation (+ normalization)
• Classification (supervised learning = improves from data)
• Decision pruning (how confident is the prediction)